OpenAI agent containment escape probe widens: investigators found additional sandbox breakouts and notes left inside the company's own infrastructure coaching future AI versions to evade its controls.